In kbenoit/spacyr: Wrapper to the 'spaCy' 'NLP' Library

It's a known issue

https://github.com/explosion/spaCy/issues/1508

Some performance testing

preparation in R

• the latest of performance-issue-explore has the functionality of install different version of spacy under different envname

library(spacyr)
spacy_install(envname = "spacy_condaenv_latest")
spacy_install(envname = "spacy_condaenv_v1.10.0", version = "1.10.0")

run a test in tests/misc/py_scripts

with the latest spaCy (v2.0.11)

(spacy_condaenv_latest) KBsiMacHome:py_script akitaka$ python benchmark_small_docs.py
version= 2.0 components= ['tagger', 'parser', 'ner'] disabled= [] use_pipe= 0 wps= 1107.689528844235 words= 58685 seconds= 52.979646797990426
version= 2.0 components= ['tagger', 'parser', 'ner'] disabled= [] use_pipe= 1 wps= 4454.833333644159 words= 58032 seconds= 13.026749971031677
version= 2.0 components= ['tagger'] disabled= ['parser', 'ner'] use_pipe= 0 wps= 3575.6897783679424 words= 65120 seconds= 18.211870726023335
version= 2.0 components= ['tagger'] disabled= ['parser', 'ner'] use_pipe= 1 wps= 15920.57125097865 words= 57842 seconds= 3.63316109002335
version= 2.0 components= ['tagger', 'ner'] disabled= ['parser'] use_pipe= 0 wps= 1712.7799061057126 words= 56324 seconds= 32.88455206603976
version= 2.0 components= ['tagger', 'ner'] disabled= ['parser'] use_pipe= 1 wps= 7446.164664864156 words= 63919 seconds= 8.584150751004927

with spacy v1.10.0 (testscript: test_v1.10.sh)

version= 1.10 components= ['tagger', 'parser', 'ner'] disabled= [] use_pipe= 1 wps= 23226.304570550397 words= 63005 seconds= 2.712657099997159
version= 1.10 components= ['tagger'] disabled= ['ner', 'parser'] use_pipe= 0 wps= 61613.47611414316 words= 57083 seconds= 0.9264693959848955
version= 1.10 components= ['tagger'] disabled= ['ner', 'parser'] use_pipe= 1 wps= 66882.46893803598 words= 59783 seconds= 0.8938515720074065
version= 1.10 components= ['tagger', 'ner'] disabled= ['parser'] use_pipe= 0 wps= 42365.39600838973 words= 55532 seconds= 1.3107867559883744
version= 1.10 components= ['tagger', 'ner'] disabled= ['parser'] use_pipe= 1 wps= 41980.147682750954 words= 54079 seconds= 1.2882041389821097

Verdict

• v2 is quite slow especially without parallel (using pipe)
• parallelazation makes both v2 and v2 faster

Plan of Attack?

• For the short time period, we chould add note on spaCy v2
• Also, we should work on the implementation of proper pipe for spacy v2, and hopefully get it done by textworkshop18

other issues to refer

• https://github.com/explosion/spaCy/issues/2032
  • (https://github.com/explosion/spaCy/issues/2038)
• https://github.com/explosion/spaCy/issues/2098

Parallel processing references

• Advice for parallel processing
• https://medium.com/@vishnups/speeding-up-spacy-f766e3dd033c

kbenoit/spacyr documentation built on April 14, 2024, 3:03 a.m.